Diagnosis of head and neck cancer by AI-based tumor-educated platelet RNA profiling of liquid biopsies

Abstract

Over 95% of head and neck cancers are squamous cell carcinoma (HNSCC). HNSCC is mostly diagnosed late, causing a poor prognosis despite the application of invasive treatment protocols. Tumor-educated platelets (TEPs) have been shown to hold promise as a molecular tool for early cancer diagnosis. We sequenced platelet mRNA isolated from blood of 101 patients with HNSCC and 101 propensity-score matched noncancer controls. Two independent machine learning classification strategies were employed using a training and validation approach to identify a cancer predictor: a particle swarm optimized support vector machine (PSO-SVM) and a least absolute shrinkage and selection operator (LASSO) logistic regression model. The best performing PSO-SVM predictor consisted of 245 platelet transcripts and reached a maximum area under the curve (AUC) of 0.87. For the LASSO-based prediction model, 1,198 mRNAs were selected, resulting in a median AUC of 0.84, independent of HPV status. Our data show that TEP RNA classification by different AI tools is promising in the diagnosis of HNSCC.